Architettura ALMA

Questi contenuti non sono ancora disponibili nella tua lingua.

Architettura ALMA

Il Triangolo IN-1

      CONVERSATORE (Haiku 4.5)
            ↓
      ANALIZZATORE (Sonnet 3.5)
            ↓
         POETA (Sonnet 3.5)

Modelli e Parametri

Componente	Modello	Temperature	Max Tokens	Scopo
Conversatore	`claude-haiku-4.5`	0.4	200	Presenza conversazionale
Analizzatore ALMA	`claude-3-5-sonnet`	0.3	600	Analisi emotiva profonda
Poeta	`claude-3-5-sonnet`	0.7	150	Generazione poetica

Perché questi modelli?

Haiku 4.5 per conversazione: veloce, economico, empatico
Sonnet 3.5 per analisi: profondità emotiva, riconoscimento shadow/light
Sonnet 3.5 per poesia: creatività, temperature più alta per improvvisazione

Additional Parameters (all calls):

top_p: 0.9
API: OpenRouter (https://openrouter.ai/api/v1/chat/completions)
Headers: HTTP-Referer, X-Title, Authorization

Flusso Completo

Step 1: Safety Check

// emotionAnalyzer.js - checkSafety()
const SAFETY_KEYWORDS = [
  'suicid', 'uccid', 'ammazzar', 'farla finita',
  'autolesion', 'tagliar', 'morti',
  'abuse', 'violen', 'picchia'
];

Se trovata keyword critica → messaggio supporto + helpline.

Step 2: Conversation Loop (MAX 5 turni)

// in1.js - Line 197-212
const response = await fetch('https://openrouter.ai/api/v1/chat/completions', {
  body: JSON.stringify({
    model: 'anthropic/claude-haiku-4.5',
    messages: [
      { role: 'system', content: CONVERSATION_PROMPT },
      ...conversationHistory
    ],
    temperature: 0.4,
    max_tokens: 200,
    top_p: 0.9
  })
});

Durante ogni turno:

Haiku 4.5 risponde con presenza (no giudizio, no consigli)
detectEmotion(userMessage) analizza ogni messaggio utente (V1 pattern-matching)
Emotion buffer accumula emozioni rilevate

Trigger per uscita:

Haiku risponde con [READY_FOR_DONO] (sente connessione)
Oppure raggiunto MAX_TURNS: 5

Step 3: ALMA V2 Analyzer

// in1.js - Line 256-293
const emotionData = await analyzeEmotionWithRetry(messages, apiKey, 2);

async function analyzeEmotion(messages, apiKey) {
  const response = await fetch('https://openrouter.ai/api/v1/chat/completions', {
    body: JSON.stringify({
      model: 'anthropic/claude-3-5-sonnet',
      messages: [
        { role: 'system', content: EMOTION_ANALYZER_PROMPT },
        ...messages
      ],
      temperature: 0.3,
      max_tokens: 600,
      top_p: 0.9
    })
  });

  return JSON.parse(response.content); // Returns emotion analysis
}

Output ALMA V2:

{
  "emotion": "tristezza",
  "intensity": 8,
  "shadow": "l'assenza che pesa",
  "light": "la capacità di ricordare",
  "reframe": "Anche il vuoto è una forma",
  "nature_raw": ["nebbia fitta", "ramo nudo", "acqua ferma"],
  "nature_transformed": ["nebbia che si solleva", "ramo in attesa", "acqua che riflette"]
}

Retry logic: 2 tentativi con 500ms wait tra essi.

Fallback: Se ALMA V2 fallisce → usa emotion buffer V1 + translateToNature()

Step 4: Poeta (MODEL B)

// almaRhama.js - generateBashoPoetry()
const response = await fetch('https://openrouter.ai/api/v1/chat/completions', {
  body: JSON.stringify({
    model: 'anthropic/claude-3-5-sonnet',
    messages: [
      { role: 'system', content: BASHO_POET_PROMPT_IT },
      {
        role: 'user',
        content: `Natura grezza: ${natureRaw}
Natura trasformata: ${natureTransformed}

Reframe: "${emotionData.reframe}"
Luce nascosta: ${emotionData.light}

Componi 3 righe:`
      }
    ],
    temperature: 0.7,
    max_tokens: 150,
    top_p: 0.9
  })
});

Output: 3 righe di poesia moderna italiana/inglese.

Step 5: Cleanup

// Remove dashes, markdown, extra whitespace
rhama = rhama.replace(/^[—\-]\s*/gm, '');
rhama = rhama.replace(/\*\*/g, '');
rhama = rhama.trim();

Step 6: Save to Database

// in1.js - Line 286-290
const insertResult = await pool.query(`
  INSERT INTO in1.rhamas (
    rhama_text,
    original_language,
    translations,
    created_at
  )
  VALUES ($1, $2, $3::jsonb, NOW())
  RETURNING id
`, [rhama, userLang, JSON.stringify(translations)]);

Step 7: Display to User

Frontend riceve Rhama + loader morphing animation.

Fallback Chain

Scenario 1: ALMA V2 Analyzer Fails

ALMA V2 FAILS (timeout/error)
    ↓
Use emotion buffer (detectEmotion results)
    ↓
Find dominant emotion (highest intensity)
    ↓
translateToNature(emotion, intensity)
    ↓
Call Poet with V1 format (nature array only)
    ↓
Generate Rhama

Scenario 2: No Emotion Detected

No clear emotion in buffer
    ↓
Use NATURE_MAP['neutro']
    ↓
Generate Rhama with neutral elements

Scenario 3: Gentle End (No Rhama)

After 5 turns, user not engaged
    ↓
Skip ALMA analyzer
    ↓
Return gentle closing message
    ↓
No Rhama generated (respects user's space)

File Locations

Backend Core

E:\BLACKTRAILS-PLATFORM\src\public\routes\in1.js
  - Main API endpoint: POST /api/rhama
  - Lines 131-430: Dual-LLM orchestration
  - Lines 197-212: MODEL A call
  - Lines 256-293: ALMA V2 analyzer call
  - Lines 266-355: Fallback logic
  - Lines 286-290: Database save

E:\BLACKTRAILS-PLATFORM\src\public\utils\almaRhama.js
  - Lines 10-11: MODEL_A, MODEL_B constants
  - Lines 17-46: NATURE_MAP
  - Lines 51-104: CONVERSATION_PROMPT
  - Lines 106-194: BASHO_POET_PROMPT_IT
  - Lines 196-284: BASHO_POET_PROMPT_EN
  - Lines 292-305: detectLanguage()
  - Lines 315-364: detectEmotion()
  - Lines 373-382: translateToNature()
  - Lines 390-393: extractReply()
  - Lines 415-486: generateBashoPoetry()

E:\BLACKTRAILS-PLATFORM\src\public\utils\emotionAnalyzer.js
  - Lines 10-22: SAFETY_KEYWORDS
  - Lines 24-33: SAFETY_RESPONSE
  - Lines 35-46: checkSafety()
  - Lines 52-196: EMOTION_ANALYZER_PROMPT
  - Lines 204-256: analyzeEmotion()
  - Lines 265-287: analyzeEmotionWithRetry()

Frontend

E:\BLACKTRAILS-PLATFORM\src\public\public\js\in1\rhama.js
  - Lines 80-110: callRhamaAPI()
  - Lines 116-178: startExperience()
  - Lines 180-253: handleChatSubmit()
  - Lines 42-74: Session recovery (F5 protection)

E:\BLACKTRAILS-PLATFORM\src\public\public\js\in1\chat-ui.js
  - Lines 42-108: Loader morphing system
  - Lines 141-187: typeWriterAI()
  - Lines 233-284: showRhamaFinale()

Session State Management

Session Storage (F5 Recovery)

sessionStorage.setItem('in1-session', JSON.stringify({
  conversationHistory,
  userEmotions,
  turnCount,
  hasRhama,
  timestamp: Date.now()
}));

Recovery: Se utente preme F5, sessione viene ripristinata da sessionStorage.

Performance Considerations

LLM Calls per Session

Normal flow: 3 calls (Conversation → Analyzer → Poet)
Fallback flow: 2 calls (Conversation → Poet with V1 emotions)
Gentle end: 1 call (Conversation only)

Cost Optimization

Haiku 4.5 per conversazione (cheap, fast)
Sonnet 3.5 solo per analisi profonda e poesia (expensive, slow)
Max tokens limitati (200/600/150) per ridurre costi

Timeout Handling

Analyzer: 2 retry attempts with 500ms wait
Frontend: 30s timeout per API call
Graceful fallback to V1 emotions if analyzer fails

Security

Safety Layer

Pre-conversation check: Crisis keyword detection
No prompt injection: System prompts isolated from user input
User isolation: Database queries scoped to user_id
Rate limiting: (TODO) 5 sessions per hour per IP

Data Privacy

Conversations not stored in database (only final Rhama)
Session storage cleared after 24h inactivity
No PII collected beyond conversation context

Questa architettura rappresenta 6+ mesi di iterazioni, fallimenti, e scoperte. ALMA non è un chatbot - è una presenza che trasforma il dialogo in natura incarnata. 🌲